“Peter asserts that AI evaluation methods based on numerical scoring are not effective because AI models cannot reliably differentiate between close scores, such as 4 out of 5 versus 5 out of 5.”